中国人工智慧新创公司 DeepSeek 于周末发布了最新模型更新 V4.1-Flash-1017，目前已在网页端、应用程式及开发者平台上线。官方表示，这次针对先前 V4.1-Flash 的小幅升级具备更强的基准测试表现、更快的任务完成速度，以及更精简的输出内容，进一步强化其对抗美国矽谷巨头 Anthropic、OpenAI 与中国本土竞争对手如 Moonshot Kimi 和 Z.ai 的实力。
根据该杭州新创公司在 Hugging Face 发布的技术报告与 X 平台贴文，新模型在推理、程式编写、代理任务和视觉基准测试中皆有显著提升，并强化了从游戏开发到 3D 建模等多元应用的能力。在衡量数学、科学与人文学科表现的「人类终极考试」（HLE）基准测试中，新版得分提升至 36.8 到 42.6 分，优于前代 V4.1-Flash 的 36.8 到 39.1 分，凸显出中国开源权重模型与高成本美国专有模型之间日益激烈的竞争。
此次更新延续了 DeepSeek 九月初推出精简版 V4.1-Flash 时的低价策略，当时的大幅折扣曾导致 MiniMax、Z.ai 甚至阿里巴巴等竞争对手的香港股价重挫。然而，极具吸引力的定价也引发大量开发者涌入并对伺服器造成沉重压力，导致 DeepSeek 近期服务减缓与中断频传，本周稍早更发生长达约五小时的当机事件，部分专家认为这归咎于运算资源不足。
Chinese AI startup DeepSeek released its latest model update, the V4.1-Flash-1017, over the weekend, making it live across the web, its app, and its developer platform. The company stated that this minor upgrade to the previous V4.1-Flash features stronger benchmark performance, faster task completion, and shorter outputs, ramping up its challenge to Silicon Valley giants like Anthropic and OpenAI as well as domestic rivals such as Moonshot Kimi and Z.ai.
According to the Hangzhou-based startup's technical report on Hugging Face and posts on X, the new model delivers substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks, while enhancing capabilities in applications ranging from game creation to 3D modeling. On the Humanity's Last Exam (HLE) benchmark, which evaluates performance across math, sciences, and humanities, the revised model scored between 36.8 and 42.6, surpassing its predecessor's 36.8 to 39.1 and underscoring the intensifying contest between China's open-weight models and high-cost US proprietary models.
This update follows the aggressive pricing strategy of the slimmed-down V4.1-Flash unveiled in early September, which offered such a steep discount that shares of rivals like MiniMax, Z.ai, and Alibaba plummeted on the Hong Kong stock exchange. However, strong developer demand for its low-priced models has strained DeepSeek's services and led to increased slowdowns and disruptions in recent months, including a five-hour outage earlier this week that some experts attributed to insufficient computing resources.